Arabic Sentiment Analysis with Federated Deep Learning

Abstract

The application of deep learning techniques in federated learning environments has shown remarkable performance across various domains. This has enabled the development of large-scale systems that enhance responsiveness, reduce processing costs and complexity, and maintain data privacy. In this research paper, we propose a federated deep learning model specifically designed for Arabic sentiment analysis using a Twitter-based benchmark dataset. Our approach leverages the effectiveness of fine-tuning the BERT model as a global learning model to extract discriminating embeddings from Arabic tweets. Through the efficient federated environment, we successfully learn the text patterns and train a classifier with the ability to accurately categorize tweets as positive, negative, or neutral. Despite the inherent complexity of Arabic language processing, extensive experiments were conducted to evaluate the performance of the federated approach in Arabic sentiment analysis. The results demonstrated significant advantages over centralized learning, particularly in terms of training time. Furthermore, our proposed model achieved a weighted average accuracy of 90% across various training and aggregation setups.